I spent GH₵142 testing four models on 120 Jumia listings and learned which copy makes Accra shoppers click
I sell home goods and small electronics on Jumia—bluetooth speakers, LED desk lamps, throw pillows, non-stick frying pans, phone cases—and I was rewriting the same feature list over and over. "Durable material. Long battery life. Available in three colours." My click-through rate from search sat at 1.8 per cent, which Jumia's seller dashboard told me was below category average. I knew the photos were fine; I suspected the descriptions were putting people to sleep.
In late January I ran a test. I took thirty products in each of four categories—electronics, fashion, home, beauty—and wrote new descriptions with four different AI models: Claude Sonnet 4.5, Gemini Flash, Llama 3.3 70B, and Mistral Large. I gave each model the same prompt, rotated which model wrote which product so no category was stuck with one voice, published all 120 new listings on the same day, and tracked clicks and add-to-carts for three weeks. Accra and Kumasi shoppers saw the new copy; I watched what they did.
One model wrote descriptions that drove 3.4 per cent click-through and converted at checkout. Another buried the benefit under two paragraphs of specs and lost people before they scrolled. Here's what I learned, the prompt I used, and the cost breakdown in cedis.
1. Claude Sonnet 4.5 wrote the tightest mobile-first copy and won on clicks
Claude's descriptions were short, benefit-first, and formatted for a phone screen. For a GH₵89 bluetooth speaker I sold in three colours, it wrote: "Pairs with your phone in under five seconds. Eight-hour battery lasts through a full day at the shop or a weekend trip to Kokrobite. Fits in a backpack side pocket. Water-resistant if it rains." Four sentences, each one a reason to buy, no filler. The product pulled a 3.4 per cent click-through rate and a 9.2 per cent add-to-cart rate from clicks—both well above my old average.
Claude understood that Jumia shoppers in Accra are scrolling on mobile, often on slower connections, and they're deciding in seconds. It front-loaded the benefit, kept paragraphs to two lines, and used specifics: "eight-hour battery" beats "long-lasting battery", "Kokrobite" beats "the beach". The model also wrote naturally for Ghanaian buyers without me having to specify it—phrases like "at the shop" and "if it rains" felt local, not translated.
Cost: GH₵0.52 per description at current cedi rates, so GH₵62.40 for 120 products. I used Kryotta's workspace to batch the prompts and export the copy into a spreadsheet, then uploaded to Jumia's bulk editor.
2. Gemini Flash was fast and cheap but wrote like a spec sheet, and mobile-money checkout suffered
Gemini Flash cost me GH₵18 for the same 120 descriptions—about one-third of Claude's price—but the output read like a user manual. For the same bluetooth speaker, it wrote: "This bluetooth speaker features bluetooth 5.0 connectivity technology, providing stable wireless transmission up to ten metres. The built-in rechargeable lithium battery offers up to eight hours of continuous playback time. The device is rated IPX5 for water resistance and includes a 3.5mm auxiliary input for wired connections." Accurate, but nobody cares about "bluetooth 5.0 connectivity technology" when they're trying to decide if this speaker is worth GH₵89 before their MoMo balance runs low.
The click-through rate for Gemini's products was 2.1 per cent—barely better than my old copy—and the add-to-cart rate was 5.8 per cent. I think people clicked because the photos were good, then bounced when the description didn't give them a reason to commit. On mobile, every extra sentence is a chance to lose someone, and Gemini gave me six sentences where Claude gave me four.
If you're selling on WhatsApp and you need 200 product descriptions drafted in an hour so you can edit them yourself, Gemini Flash is fine. For Jumia listings where the copy has to close the sale on its own, it's not worth the savings.
3. Llama 3.3 70B wrote the most creative descriptions but over-explained features that didn't matter
Llama surprised me. It wrote descriptions that felt like a person was talking to you, with personality and detail, but it often picked the wrong detail. For a set of non-stick frying pans (GH₵135 for three sizes), it wrote: "The aluminium core heats evenly so your kontomire doesn't burn on one side while the other stays raw. The non-stick coating means you use less oil, which saves money if you're cooking every day. The handles stay cool even when the pan's been on the fire for twenty minutes, so you won't burn your hand reaching for the spatula."
I loved the kontomire reference—it's specific, it's real—but the description kept going for another three sentences about the rivets and the pan's weight distribution. By the time someone reading on a phone got to the price, they'd forgotten why they cared. Click-through was 2.6 per cent, add-to-cart was 6.4 per cent. Better than Gemini, worse than Claude.
Llama's strength is voice. If you're writing Instagram captions or WhatsApp broadcast messages where personality matters more than brevity, it's excellent. For Jumia product descriptions where you have five seconds to make the case, it needs a tighter leash.
4. Mistral Large wrote the longest descriptions and lost people halfway through
Mistral's output was detailed, structured, and too long for mobile. For a GH₵45 phone case, it wrote a description that included four subheadings: "Protection Features", "Design & Aesthetics", "Compatibility", and "Value Proposition". Each section was two to three sentences. The total word count was 140 words. On a phone screen, that's a lot of scrolling for a product that costs less than a week's tro-tro fare.
Click-through: 1.9 per cent. Add-to-cart: 5.1 per cent. People saw the wall of text and kept scrolling. I don't think they even read it.
Mistral might work for high-ticket items where buyers expect detail—maybe a GH₵3,500 laptop or a GH₵1,200 sewing machine—but for everyday products on Jumia, it's overkill. The model assumes people have time and patience, and Accra shoppers scrolling on their lunch break have neither.
5. The prompt template that worked across all four models
I used the same prompt for each model, which made the test fair and also taught me what each one does with the same instructions. Here's the template:
"Write a product description for [product name] that will appear on a Jumia listing in Ghana. The price is [price in cedis]. Focus on the main benefit in the first sentence, then give two or three specific reasons a shopper in Accra or Kumasi would choose this product. Keep it under 80 words. Write for someone reading on a phone who's deciding quickly. Don't use phrases like 'premium quality' or 'perfect for'—be specific instead."
I filled in the product name, price, and any key detail (colour options, size, battery life), then ran the prompt. Claude and Llama followed the 80-word limit; Gemini came in around 90; Mistral ignored it and wrote 120+. That told me something about how each model weights instructions versus its own defaults.
If you're batching prompts for 50 or 100 products, save the template as a snippet in Kryotta's workspace and swap in the variables. It took me about ninety minutes to generate all 120 descriptions across four models, including the time to export and format for Jumia's bulk uploader.
6. Cost breakdown: GH₵142 total, Claude worth the premium for listings that convert
Here's what I spent on 120 descriptions (30 per model):
- Claude Sonnet 4.5: GH₵62.40 (GH₵0.52 per description)
- Gemini Flash: GH₵18.00 (GH₵0.15 per description)
- Llama 3.3 70B: GH₵34.80 (GH₵0.29 per description)
- Mistral Large: GH₵26.80 (GH₵0.22 per description)
Total: GH₵142.00
Claude cost more than three times what Gemini did, but the click-through rate was 62 per cent higher and the add-to-cart rate was 59 per cent higher. If you're selling a product with a GH₵50 margin and Claude's copy gets you five extra sales per month, the model pays for itself in the first week.
I'm now using Claude for all new Jumia listings and for any product that's underperforming. I'll use Gemini Flash for quick Instagram captions or WhatsApp broadcasts where I'm going to edit the output anyway, but not for copy that has to convert on its own.
7. What didn't work: trying to make one model sound like another
Halfway through the test I tried to "fix" Gemini by adding "write in a conversational tone, like you're texting a friend" to the prompt. It helped a little—fewer phrases like "features advanced technology"—but it still wrote longer and more formal than Claude. I also tried to rein in Mistral by saying "maximum 60 words, hard limit", and it gave me 95.
Each model has a default voice and structure, and you can nudge it but you can't rewrite its personality. It's faster to pick the model that naturally does what you need than to spend twenty minutes editing the output of a cheaper one.
8. One surprise: Claude wrote better fashion descriptions than I expected
I thought Claude would excel at electronics and specs-heavy products, but the fashion descriptions were the most impressive. For a GH₵68 ankara print dress, it wrote: "The print stays bright after washing. The waist tie adjusts so it fits whether you're seven months along or back to your usual size. Pockets deep enough for your phone and keys." Three benefits, each one concrete, no fluff about "elegant style" or "timeless design". The dress sold out in eleven days.
Fashion is hard to write for because the benefits are subjective—comfort, fit, style—but Claude found specific, believable claims. I think it's because the model is good at inferring what matters to the buyer (pockets, adjustable fit, colour that lasts) instead of describing the garment in abstract terms.
Questions people ask
Which model should I use if I'm writing 200 Jumia descriptions in one sitting?
Claude Sonnet 4.5 if you want copy that converts and you're willing to spend GH₵104 for 200 products. Gemini Flash if you need a rough draft fast and you'll edit every one yourself—it'll cost you GH₵30 but plan on two hours of rewriting.
Can I use the same AI-generated description for Jumia and Instagram?
Not without editing. Jumia descriptions need to be benefit-first and scannable; Instagram captions need personality and a hook in the first line. I use Claude for Jumia, Llama for Instagram, and I rewrite the first sentence of each Instagram caption by hand.
Does Jumia's algorithm favour certain words or length in product descriptions?
Jumia's seller documentation says the algorithm prioritises click-through rate and conversion rate, not keyword density or word count. I didn't see any evidence that longer descriptions ranked better. The products with Claude's short, tight copy got more visibility in search results, probably because they converted better and the algorithm rewarded that.
How often should I rewrite product descriptions that aren't selling?
If a product's been live for two weeks with fewer than twenty views, rewrite the title first—that's what shows up in search—then the description. I rewrite descriptions every month for my bottom ten products by views. It takes about thirty minutes with Claude and it's pulled three products back from "dead stock" status in the last two months.
I've kept Claude as my default for Jumia product descriptions, and my overall click-through rate is now 3.1 per cent across 240 active listings. The time I save not rewriting the same feature list over and over is worth more than the GH₵0.52 per product, and the conversion lift paid for the test cost in the first week. If you're selling on Jumia and your descriptions still sound like a spec sheet, try Claude with the prompt template above—you can test it on ten products in Kryotta's workspace and see the difference before you commit to a full rewrite.
Start testing AI models for your Jumia listings at Kryotta—Claude, Gemini, Llama and Mistral in one workspace, with side-by-side comparison and export to spreadsheet.
